以下文案源自于微X公众号：算法美食屋
作者：梁云1991
链接：https://mp.weixin.qq.com/s/KtXsPjmWhOpylO1EDdM7Fg
本文仅用于学术分享，如有侵权，请联系后台作删文处理  导读在模型训练过程中，常常会打印每一个epoch训练的日志，令人眼花缭乱。本文运用torchkeras库中的VLog，分别仔细的演示了在LightGBM、ultralytics、transformers中运用VLog，并供给了源码，期盼对大众有帮忙。训练日志刷屏使我痛苦，我研发了VLog，能够在任意训练代码中容易运用~例如，经过callback嵌入到lightgbm/catboost/transformers/ultralytics，乃至keras库的训练代码流程中~
before:
after：
为何不消tensorboard或wandb? tensorboard需要开端口权限，服务器研发环境有时候无端口权限~wandb需要联网，有时候网速很差或无网，影响体验~综合对比思虑如下表
一，VLog基本原理VLog类重点有以下5个办法。from torchkeras import
VLog #1, 初始化办法vlog = VLog(epochs=20, monitor_metric=val_loss, monitor_mode=min
) #2, 表示起始空图表
vlog.log_start() #3, 更新step级别日志vlog.log_step({train_loss:0.003,val_loss:0.002
}) #4, 更新epoch级别日志vlog.log_epoch({train_acc:0.9,val_acc:0.87,train_loss:0.002,val_loss:0.03
}) #5, 输出最后稳定状态图表
vlog.log_end() import
time import
math,random from torchkeras import
VLog epochs = 10batchs = 30#0, 指定监控北极星指标，以及指标优化方向vlog = VLog(epochs, monitor_metric=val_loss, monitor_mode=min
) #1, log_start 初始化动态图表
vlog.log_start() for epoch in
range(epochs): #train for step in
range(batchs): #2, log_step 更新step级别日志信息，打日志，并用小进度条表示进度 vlog.log_step({train_loss:100-2.5*epoch+math.sin(2
*step/batchs)}) time.sleep(0.05
) #eval for step in range(20
): #3, log_step 更新step级别日志信息，指定training=False说明在验证模式，只打日志不更新小进度条 vlog.log_step({val_loss:100-2*epoch+math.sin(2*step/batchs)},training=False
) time.sleep(0.05
) #4, log_epoch 更新epoch级别日志信息，每一个epoch刷新一次动态图表和大进度条进度 vlog.log_epoch({val_loss:100 - 2*epoch+2*random.random()-1
, train_loss:100-2.5*epoch+2*random.random()-1
}) # 5, log_end 调节坐标轴范围，输出最后指标可视化图表
vlog.log_end()
二，在LightGBM中运用VLog
设计一个简单的回调，就能够搞定~ from torchkeras import
VLog class VLogCallback: def __init__
(self, num_boost_round, monitor_metric=val_loss
, monitor_mode=min)
: self.order = 20
self.num_boost_round = num_boost_round
self.vlog = VLog(epochs = num_boost_round, monitor_metric = monitor_metric,
monitor_mode = monitor_mode) def __call__(self, env) -> None:
metrics = {} for item in
env.evaluation_result_list: if len(item) == 4
: data_name, eval_name, result = item[:3
] metrics[data_name+_
+eval_name] = result else
: data_name, eval_name = item[1
].split() res_mean = item[2
] res_stdv = item[4
] metrics[data_name+_
+eval_name] = res_mean
self.vlog.log_epoch(metrics) import
datetime import numpy as
np import pandas as
pd import lightgbm as
lgb from sklearn import
datasets fromsklearn.model_selectionimport
train_test_split from sklearn.metrics import
accuracy_score def printlog(info): nowtime = datetime.datetime.now().strftime(%Y-%m-%d %H:%M:%S
) print("\n"+"=========="*8 + "%s"
%nowtime) print(info+...\n\n
) #================================================================================# 一，读取数据#================================================================================printlog("step1: reading data..."
) # 读取dftrain,dftest
breast = datasets.load_breast_cancer() df = pd.DataFrame(breast.data,columns = [x.replace(,_) for x in
breast.feature_names]) df[label
] = breast.target df[mean_radius] = df[mean_radius].apply(lambda
x:int(x)) df[mean_texture] = df[mean_texture].apply(lambda
x:int(x))
dftrain,dftest = train_test_split(df) categorical_features = [mean_radius,mean_texture
] lgb_train = lgb.Dataset(dftrain.drop([label],axis = 1),label=dftrain[label
],
categorical_feature = categorical_features) lgb_valid = lgb.Dataset(dftest.drop([label],axis = 1),label=dftest[label
],
categorical_feature = categorical_features,
reference=lgb_train) #================================================================================# 二，设置参数#================================================================================printlog("step2: setting parameters..."
) boost_round = 50early_stop_rounds = 10
params = { boosting_type: gbdt
, objective:binary
, metric: [auc], #l2 num_leaves:15
, learning_rate: 0.05
, feature_fraction: 0.9
, bagging_fraction: 0.8
, bagging_freq: 5
, verbose: 0
, early_stopping_round:5
} #================================================================================# 三，训练模型#================================================================================printlog("step3: training model..."
)
result = {} vlog_cb = VLogCallback(boost_round, monitor_metric = val_auc, monitor_mode = max
)
vlog_cb.vlog.log_start()
gbm = lgb.train(params,
lgb_train,
num_boost_round= boost_round,
valid_sets=(lgb_valid, lgb_train), valid_names=(val,train
),
callbacks = [lgb.record_evaluation(result),
vlog_cb]
)
vlog_cb.vlog.log_end() #================================================================================# 四，评定模型#================================================================================printlog("step4: evaluating model ..."
) y_pred_train = gbm.predict(dftrain.drop(label,axis = 1
), num_iteration=gbm.best_iteration) y_pred_test = gbm.predict(dftest.drop(label,axis = 1
), num_iteration=gbm.best_iteration) print(train accuracy: {:.5} .format(accuracy_score(dftrain[label],y_pred_train>0.5
))) print(valid accuracy: {:.5} \n.format(accuracy_score(dftest[label],y_pred_test>0.5
))) #================================================================================# 五，保留模型#================================================================================printlog("step5: saving model ..."
) model_dir = "gbm.model"print("model_dir: %s"
%model_dir) gbm.save_model("gbm.model"
) printlog("task end..."
) ######
================================================================================2023-11-10 15:39:38
step1: reading data......
================================================================================2023-11-10 15:39:38
step2: setting parameters......
================================================================================2023-11-10 15:39:38
step3: training model......
================================================================================2023-11-10 15:39:44
step4: evaluating model ......
train accuracy: 0.95775
valid accuracy: 0.94406
================================================================================2023-11-10 15:39:44
step5: saving model ......
model_dir: gbm.model
================================================================================2023-11-10 15:39:44
task end......
三， 在ultralytics中运用VLog写个适配的回调~ultralytics能够做 归类，检测，分割 等等。这个回调函数是通用的，此处以归类问题为例，改个monitor_metric就可~cats_vs_dogs数据集能够在公众号江大白后台回复:torchkeras 获取~from torchkeras import
VLog class VLogCallback: def __init__(self,epochs,monitor_metric,monitor_mode):
self.vlog = VLog(epochs,monitor_metric,monitor_mode) def on_train_batch_end(self,trainer): self.vlog.log_step(trainer.label_loss_items(trainer.tloss, prefix=train
)) def on_fit_epoch_end(self,trainer): metrics = {k.split(/)[-1]:v for k,v in trainer.metrics.items() if loss not in
k}
self.vlog.log_epoch(metrics) def on_train_epoch_end(self,trainer): passfrom ultralytics import
YOLO epochs = 10
vlog_cb = VLogCallback(epochs = epochs, monitor_metric=accuracy_top1
, monitor_mode=max
)
callbacks = { "on_train_batch_end"
: vlog_cb.on_train_batch_end, "on_fit_epoch_end"
: vlog_cb.on_fit_epoch_end
} model = YOLO(model = yolov8n-cls.pt
) forevent,funcin
callbacks.items():
model.add_callback(event,func)
vlog_cb.vlog.log_start() results = model.train(data=cats_vs_dogs
, epochs=epochs, workers=4) # train the model
vlog_cb.vlog.log_end()
四， 在transformers中运用VLogwaimai评论数据集能够在公众号江大白后台回复:torchkeras 获取~#回调给你写好了~from torchkeras.tools.transformers import
VLogCallback import numpy as
np import pandas as
pd import
torch import
datasets from transformers import
AutoTokenizer,DataCollatorWithPadding from transformers import
AutoModelForSequenceClassification from transformers import
TrainingArguments,Trainer from transformers import
EarlyStoppingCallback from tqdm import
tqdm from transformers import
AdamW, get_scheduler #一，准备数据df = pd.read_csv("waimai_10k.csv"
)
ds = datasets.Dataset.from_pandas(df) ds = ds.shuffle(42
) ds = ds.rename_columns({"review":"text","label":"labels"
}) tokenizer = AutoTokenizer.from_pretrained(bert-base-chinese
) ds_encoded = ds.map(lambda example:tokenizer(example["text"
]), remove_columns = ["text"
], batched=True
) #train,val,test splitds_train_val,ds_test = ds_encoded.train_test_split(test_size=0.2
).values() ds_train,ds_val = ds_train_val.train_test_split(test_size=0.2
).values()
data_collator = DataCollatorWithPadding(tokenizer=tokenizer) dl_train = torch.utils.data.DataLoader(ds_train, batch_size=16
, collate_fn = data_collator) dl_val = torch.utils.data.DataLoader(ds_val, batch_size=16
, collate_fn = data_collator) dl_test = torch.utils.data.DataLoader(ds_test, batch_size=16
, collate_fn = data_collator) for batch in
dl_train: breakprint({k: v.shape for k, v in
batch.items()}) #二，定义模型
model = AutoModelForSequenceClassification.from_pretrained( bert-base-chinese,num_labels=2
) #三，训练模型def compute_metrics(eval_preds):
logits, labels = eval_preds preds = np.argmax(logits, axis=-1
)
accuracy = np.sum(preds==labels)/len(labels) precision = np.sum((preds==1)&(labels==1))/np.sum(preds==1
) recall = np.sum((preds==1)&(labels==1))/np.sum(labels==1
) f1 = 2
*recall*precision/(recall+precision) return {"accuracy":accuracy,"precision":precision,"recall":recall,f1
:f1}
training_args = TrainingArguments( output_dir = "bert_waimai"
, num_train_epochs = 3
, logging_steps = 20
, gradient_accumulation_steps = 10
, evaluation_strategy="steps", #epoch metric_for_best_model=eval_f1
, greater_is_better=True
, report_to=none
, load_best_model_at_end=True
) callbacks = [EarlyStoppingCallback(early_stopping_patience=10
), VLogCallback()] #监控指标同 metric_for_best_model
trainer = Trainer(
model,
training_args,
train_dataset=ds_train,
eval_dataset=ds_val,
compute_metrics=compute_metrics,
callbacks = callbacks,
data_collator=data_collator,
tokenizer=tokenizer,
)
trainer.train() #四，评定模型
trainer.evaluate(ds_val) #五，运用模型from transformers import
pipeline model.config.id2label = {0:"差评",1:"好评"
} classifier = pipeline(task="text-classification"
,tokenizer = tokenizer,model=model.cpu()) classifier("挺好吃的哦"
) #六，保留模型model.save_pretrained("waimai_10k_bert"
) tokenizer.save_pretrained("waimai_10k_bert"
) classifier = pipeline("text-classification",model="waimai_10k_bert"
) classifier(["味道还不错，下次再来","我去，吃了我吐了三天"
])
 fill=%23FFFFFF%3E%3Crect x=249 y=126 width=1 height=1%3E%3C/rect%3E%3C/g%3E%3C/g%3E%3C/svg%3E)
万水千山总是情，点个赞赞行不行~ |
发表于 2024-10-8 14:44:07